I Reverse-Engineered 200 AI Startups. 146 Are Selling You Repackaged ChatGPT.
Source: https://www.techcrunch.com/2025/11/25/… Author: Teja Kusireddy Date: 2025-11-25
Summary
Analysis of 200 AI startups finding that the majority are thin wrappers on top of foundation models with no real moat. Categorizes which have genuine defensibility and which don’t.
Key Claims
- 73% of surveyed AI startups: “wrapper” companies — a system prompt + basic UI on top of an OpenAI/Anthropic API.
- The wrapper problem: wrappers can be replicated in a weekend. They lose when the foundation model company adds their feature natively.
- Defensible categories: companies with proprietary data (e.g., medical, legal, industrial), companies with network effects, companies that have trained specialized fine-tuned models.
- The market dynamic: wrappers succeed in the short term (low cost to build, high initial PMF), fail in the medium term (competition from better-positioned players and model companies themselves).
Concepts
- Synthetic Data — proprietary data as the real moat
- Coding Agents — coding agent wrappers are particularly vulnerable